The components structure and principle of ring ball gratings sensor is put forward for the present state that there still has not a better method to detect torque under the extreme environment in domestic. According to the localization property in time and frequency domains and the translation invariant property, the preprocessed signal of sensor is decomposed with multi resolution, and the signal is reconstructed with wavelet under the intervention of neural network. The wavelet neural networks (WNN) model is constructed and the algorithm routine is realized in laboratory virtual instrument engineering workbench (LabVIEW). With the testing device, the demodulation experiments are carried out for the corresponding sensor signals when the shaft is continuous rotated along the same direction, rotated along the alternate direction and knocked. The results indicate that the demodulation effect and confidence level are very well.